SimpleX Tetra-1 Embeds Protein Folding Inside Its Reasoning Chain, Outpacing IsoDDE on Docking
SimpleX's biology-native model integrates protein folding natively, scoring 53.1% on Runs N' Poses

Chinese artificial intelligence startup SimpleX released a preview of Tetra-1 on November 10, 2026, describing it as a biology-native large language model that integrates protein structure prediction directly into its chain-of-thought reasoning — not by calling an external specialized model, but by running structure prediction as a native step within the same inference loop.
The claim, if it holds up under independent evaluation, would distinguish Tetra-1 architecturally from every current competitor in the AI drug-discovery space. Tools including Claude Science from Anthropic and GPT-Rosalind from OpenAI coordinate separate structure-prediction engines through external orchestration. SimpleX says Tetra-1 does not.
The practical significance of that architectural distinction rests on an unresolved question in AI drug discovery: whether a language model that reasons about molecular structure in the same computational stream as natural language can learn chemical representations that an orchestration-based system cannot. That question has driven substantial investment across the field in 2026, and Tetra-1 is the first system to claim a direct answer through published benchmark results.
Architecture: Calpha
A supporting paper posted to arXiv on November 16, 2026, titled "Calpha: Toward Native Biomolecular Intelligence in a General-Purpose Language Model", describes the integration mechanism in detail. The Calpha architecture implements a three-step loop: Tetra-1 first generates a biochemical hypothesis in natural language, then invokes SimpleFold-3B — a compact protein structure prediction model developed in collaboration with Apple Research — directly within the reasoning chain, then performs a self-check pass on the predicted structure before continuing to the next reasoning step.
The critical architectural distinction, according to the paper, is that SimpleFold-3B runs inside the LLM's inference process rather than receiving a task handoff from an orchestration layer. In an orchestration-based system, a language model calls a structure-prediction function, waits for a result, and receives that result as a text string or structured output. In Calpha, the structure prediction step is embedded as part of the same forward pass, allowing gradients and representations to flow across the boundary between language reasoning and structural computation.
The Calpha paper frames this as the difference between a tool call and a native cognitive step. Whether that description represents a meaningful engineering breakthrough or a reframing of an existing approach will depend on independent technical review of the architecture, which has not yet been conducted. The preprint was posted less than two weeks before the Tetra-1 preview release and has not yet undergone peer review.
SimpleFold-3B
SimpleFold-3B, described in a separate arXiv paper published in September 2025 and accepted at ICLR 2026, was designed from the outset for integration inside a general-purpose language model. The model uses only standard transformer blocks with adaptive layer normalization, explicitly avoiding triangular attention, multiple sequence alignment, and the domain-specific pair representations that have characterized prior state-of-the-art protein folding systems, including AlphaFold 2 and AlphaFold 3.
At three billion parameters, SimpleFold-3B was trained on approximately nine million distilled protein structures alongside entries from the Protein Data Bank. That training approach allows the model to generalize across a wide range of protein families without requiring the specialized preprocessing pipelines that larger folding systems depend on. The tradeoff is that SimpleFold-3B achieves lower absolute accuracy than AlphaFold 3 on well-characterized protein families with abundant structural data. Its architecture prioritizes generality over peak performance on known structures, which is the property that makes it compatible with native integration inside a general-purpose language model and potentially more effective on novel, low-homology targets.
Benchmark Results
SimpleX reports Tetra-1 scored 53.1% on the most challenging subset of the Runs N' Poses benchmark, a protein-ligand docking evaluation published in February 2025 by Škrinjar and colleagues. The benchmark's hardest subset restricts evaluation to protein-ligand complexes with no more than 20% sequence similarity between training and test structures — 76 complexes in total — specifically designed to test generalization beyond memorized structural patterns. Protein-ligand docking generalization is considered a key capability for drug discovery because pharmaceutically relevant targets often have no close structural analogs in existing databases.
For comparison, IsoDDE, the AI drug-design engine from Isomorphic Labs, reported approximately 50% on that same subset in its February 2026 technical report. AlphaFold 3 achieved approximately 23% on the same 0–20% subset, according to the IsoDDE technical report. If the figures are taken at face value, Tetra-1 would represent a 3.1 percentage point improvement over IsoDDE and more than double AlphaFold 3's score on the hardest portion of the benchmark.
The scores require careful interpretation. Tetra-1's 53.1% figure comes from SimpleX itself and has not been independently replicated. IsoDDE's 50% figure comes from Isomorphic Labs' own technical report. The benchmarks are identical and the evaluation methodology appears consistent; both companies used the same published test set. However, the results have not been validated in a neutral, independent head-to-head comparison.
SimpleX also reports results on FoldBench, a Nature Communications benchmark published in December 2025 covering 1,522 low-homology biological assemblies across nine prediction tasks. AlphaFold 3 scores approximately 70.87% on protein-protein interaction tasks and 47.95% on antibody-antigen tasks on FoldBench's public leaderboard, maintained at GitHub. SimpleX reports Tetra-1 improves on the protein-protein category figures, though Tetra-1 entries had not appeared on the public GitHub leaderboard as of November 2026.
Competitive Context
Tetra-1 enters a field that has drawn significant investment and concentrated research attention in 2026, particularly following the 2024 Nobel Prize in Chemistry, which recognized AlphaFold's contributions to protein structure prediction and signaled the scientific community's view that AI-driven structural biology had reached a foundational milestone.
Isomorphic Labs, the Google DeepMind spinout specializing in AI-accelerated drug discovery, closed a Series B round of $2.1 billion in May 2026, led by Thrive Capital. The company's total funding stands at approximately $2.6 billion. Its IsoDDE system is built on the AlphaFold 3 lineage and has been benchmarked across a range of protein-docking and molecular design tasks. IsoDDE uses an orchestration approach in which molecular design and structure prediction run as coordinated but separate processes.
The field has also attracted researchers formerly central to AlphaFold's development. John Jumper, who shared the 2024 Nobel Prize in Chemistry for AlphaFold and was then leading protein-structure AI research at Google DeepMind, left that role in June 2026 to join Anthropic, where he is working on AI-assisted scientific discovery. Anthropic, separately, launched Claude Science in June 2026 as a beta research workbench that uses orchestration-based integration for structure prediction tasks, allowing researchers to run protein folding as part of a broader scientific reasoning workflow. OpenAI released GPT-Rosalind in April 2026 for protein structure and molecular design tasks, though that system remains restricted to US-based organizations.
The entry of a Chinese startup into this competitive space carries additional context. SimpleX operates under Chinese corporate law and has not, as of November 2026, published data-access policies in English or disclosed its arrangements regarding government data requests. The Calpha paper and the Tetra-1 technical documentation released to date are primarily in Chinese, with an English abstract for the Calpha arXiv paper. Research organizations, healthcare companies, or pharmaceutical firms outside China considering use of Tetra-1 would typically conduct due diligence on jurisdictional data-access obligations separately from any technical evaluation.
Implications
Whether Tetra-1's native integration approach actually improves protein structure prediction quality over orchestration-based alternatives remains to be tested. The Calpha architecture is theoretically coherent — a system that reasons about molecular structure in the same processing stream as natural language may learn chemical representations that an external orchestration layer cannot access — but the benchmark advantage SimpleX claims is narrow, within normal variation for self-reported vendor scores, and covers only docking generalization tasks on a single benchmark subset.
A MIT Technology Review interview with a researcher in the field described the core challenge for any biology-native LLM: scientific reasoning requires grounding claims in physical structure, and language-only models have consistently failed at that grounding without external correction. Whether Tetra-1's native loop addresses that failure mode in practice, or simply relocates the correction step inside the model's inference process, is a question that independent replication will need to answer.
The Calpha paper is currently a preprint and has not completed peer review. SimpleX has described the November 2026 release as a preview. The company has not announced general availability, pricing, access terms, or external API documentation in English. The absence of an English-language technical interface and the preliminary state of the Calpha preprint mean that independent evaluation is likely months away at the earliest.



